Post on 08-Aug-2020
transcript
Systemic Risk: Stochastic Orders
Jan DhaeneFaculty of Business and Economics
Katholieke Universiteit Leuven
Jan.Dhaene@kuleuven.be
Roger J. A. LaevenAmsterdam School of Economics
University of Amsterdam, KU Leuven
and CentER
R.J.A.Laeven@uva.nl
Yiying ZhangSchool of Statistics and Data Science, LPMC and KLMDASR
Nankai University, Tianjin 300071, P. R. China
yyzhang@nankai.edu.cn
July 15, 2019
Abstract
There is pervasive interconnectedness within the complex system of financial insti-tutions. Due to this interconnectedness, a contagious disruptive failure of a singularplayer can cause a collapse in part of the system, with a reverberating effect onthe system and economy as a whole. A pivotal step in assessing this systemic riskis to develop concepts, tools, and techniques to provide a(n) (partial) ordering offinancial institutions or systems and the systemic risk borne and induced by them;to next provide a suitable measurement of this risk; and to finally attribute this riskequitably to the individual institutions that generate it. In this paper, we introducesome new stochastic orders related to systemic risk—the systemic contribution or-der, the systemic relevance order, and the systemic risk order—and analyze theircharacterizations and properties. This paper is the first in a triplet of papers on sys-temic risk by the same authors. In Dhaene et al. (2019a), we introduce conditionaldistortion risk measures and distortion risk contribution measures as measures ofsystemic risk. In Dhaene et al. (2019b), we develop procedures for attributing andallocating systemic risk among the players in a given risky environment.
Keywords: Systemic risk; Risk measures; Contagion; Financial crisis; Prudentialregulation; Conditional stochastic ordering; Stop-loss ordering.
JEL Classification: G21, G22, G31.
1
1 Introduction
The global financial crisis of 2007-2009 can be characterized by spill-over effects and
pronounced transmissions of a few adverse events that led to a cascading sequence of
many more adverse events with severe financial distress. In the United States in 2008, we
saw the collapse of a massive number of financial institutions and corporations because
of distressed results of subprime mortgage loans and credit default swaps. This rapidly
affected financial and investment activities that led to bank failures in many other parts of
the world and the downturn of the stock market that spiralled across several other stock
markets around the world. It ultimately threatened the collapse of the global financial
market to the extent that the system was on the brink of a meltdown. In an economy
composed of a group of financial institutions, a “systemic risk” is present when the failure
or loss of a singular or a few players in the economy threatens the security and stability
of the other players, and as a result of the system and economy as a whole. The high
level of interconnectedness of financial markets and institutions provides a fertile soil for
the contagious transmission and rapid propagation of adverse shocks leading to such a
systemic risk.
Expectedly thereafter, this drew the initiatives of several nations to cooperate and de-
velop increased prudential regulation and supervision of financial institutions. Developed
since 1974 by a committee of members representing the G-10 countries, the Basel Accords
(Basel Committee on Banking Supervision, 1988, 2006), comprising Basel I, Basel II and
most recently, Basel III, consist of a set of recommended actions for countries to regulate
their own banking industry. In a similar fashion though of a binding nature, the Sol-
vency Directives I and II codify insurance regulation and supervision for countries within
the European Union. Both sets of regulations are primarily motivated to ensure harmo-
nized solvency of individual banks and insurance companies by requiring them to hold
and maintain a prudent level of capital according to their respective individual risk pro-
files. Following the global financial crisis, financial regulators and supervisory authorities
are increasingly recognizing the presence of interconnectedness of financial institutions
in financial markets globally; see e.g. Basel Committee on Banking Supervision (2011).
This interconnectedness creates a form of systemic risk that should be accounted for in
the monitoring and supervision of financial institutions. Therefore, the risks borne and
induced by financial institutions should not be monitored in isolation. This insight is lead-
ing us in recent years to a major shift from microprudential regulation and supervision
2
of financial institutions to macroprudential regulation and supervision. Macroprudential
regulation is now the term being used to regulate and supervise financial institutions to al-
leviate the consequences of systemic risk; see e.g., the initiatives of the Financial Stability
Board (FSB), the European Systemic Risk Board (ESRB), and, of a more specific nature,
the European Insurance and Occupational Pensions Authority (EIOPA) which recently
published a series of papers1 with the aim of contributing to the debate on systemic risk
and macroprudential policy for insurers. Today we are seeing a growing need and interest
among financial regulators and supervisory authorities, not only to determine solvency
capital requirements at the micro level, but also to assess the aggregate risk in the fi-
nancial system from a macro perspective and identify Systemically Important Financial
Institutions (SIFIs) and Global or Domestic Systemically Important Insurers (G-SIIs or
D-SIIs).
When it comes to the definition of what constitutes a systemic risk, Kaufman and Scott
(2003) argued that “systemic risk is the risk or probability of breakdowns in an entire
system, as opposed to breakdowns in individual parts or components and is evidenced
by co-movements (correlation) among most or all of the parts.” Quite similar to the one
suggested by the Group of Ten (2001), Cummins and Weiss (2013) gave the following
definition: “Systemic risk is the risk that an event will trigger a loss of economic value
or confidence in a substantial segment of the financial system that is serious enough to
have significant adverse effects on the real economy with a high probability.” Thus, a
systemic risk can be viewed as the risk of individual adverse events, that trigger further
adverse events, to the financial system, and as a result to the real economy. While this
last definition does not directly allude to the idea of “interconnectedness” of the entities
in the financial market, it does imply that the system is intertwined to the effect that a
systemic event can reverberate to the entire or a substantial part of the financial system.
In this paper, we seek to develop the language of “systemic risk ordering” to bear on the
problem of assessing systemic risk. To this end, we introduce some new stochastic orders
related to systemic risk and analyze their characterizations and properties. We introduce
the “systemic contribution order” and the “systemic relevance order” to stochastically
compare the contributions and relevance of individual financial institutions within a fi-
nancial system. We also introduce the “systemic risk order” to provide a partial ordering
of the aggregate risk in financial systems. The ordered systems only differ in terms of the
1https://eiopa.europa.eu/Pages/News/Enhancing-the-current-macroprudential-framework-EIOPA-publishes-the-third-paper-of-a-series-on-systemic-risk-and-macroprude.aspx
3
dependence structure among the individual risks. In addition, it is important to suitably
measure systemic risk, and to attribute the aggregate risk in the system together with
its associated capital to the financial institutions that generate this risk. By “systemic
risk allocation”, we mean a fair and equitable subdivision of the aggregate risk capital
in the system across the constituents of the system. The objective of partially ordering
systemic risk and comparing the relevance and contributions of individual institutions in
the financial system from a macro perspective is the aim of the present paper. In a first
companion paper (Dhaene et al., 2019a), we propose conditional distortion risk measures
and distortion risk contribution measures to quantify conditional systemic risk events and
present sufficient conditions for two random vectors to be ordered in terms of the proposed
measures. In a second companion paper (Dhaene et al., 2019b), we consider the objective
of designing a mechanism to equitably subdivide the aggregate risk capital in the system
across its constituents according to the systemic risk that constituent generates.
There is a long and rich history in actuarial science and applied probability of partially
ordering univariate risks. Subsequently, the literature has focused its attention on the
analysis of multivariate stochastic orders. By contrast, conditional stochastic ordering and
its applications in insurance and finance, as considered in this paper, has seen relatively
little interest. Revealing the relevance of conditional stochastic ordering in an insurance
and financial context can be viewed as a separate contribution of this paper that is of
independent interest.
The remainder of this paper is organized as follows: In Section 2, we introduce the
setting and notation and some relevant definitions. In Section 3, we introduce the systemic
contribution order. In Section 4, we introduce the systemic relevance order. Section 5
defines the notion of systemic risk order. Conclusions are presented in Section 6.
2 Preliminaries
In this section, we present some definitions and notions used in the sequel. For ease of
presentation and readability of the paper, the definitions and related properties of the
canonical univariate and multivariate stochastic orders, of comonotonicity and of other
dependence notions used hereafter are relegated to the Appendix.
4
2.1 Setting, notation and definitions
All random variables (r.v.’s) considered hereafter are defined on the probability space
(Ω,F ,P). All expectations and density functions are assumed to be well-defined when
they appear. Two r.v.’s are identified if they are P-almost surely (P-a.s.) equal, and
we understand throughout equalities and inequalities between r.v.’s in the P-a.s. sense.
We denote by FX the cumulative distribution function (cdf) of a given r.v. X under the
reference probability measure P : FX(x) = P[X ≤ x]. We use the terms “increasing” and
“decreasing” in a non-strict sense. Furthermore, we use the notation ‘d=’ for equality in
distribution. Finally, x+ ≡ maxx, 0.Consider a market composed of n (non-collaborating, competitive) financial institu-
tions or financial conglomerates, the stochastic losses of which are represented by the
r.v.’s X1, . . . , Xn. Suppose that, by microprudential regulation, each individual financial
institution is required to hold a certain amount of microprudential risk capital equal to
Ri, i = 1, . . . , n, where the italic upper case “R” stands for “Required”. All of the results
developed in this paper can also be applied to the case of available capital, which falls in
the scope of our work in Dhaene et al. (2019b). The aggregate amount of microprudential
risk capital in the market, denoted by R, thus equals R =∑n
i=1Ri. We interpret Ri to
include both the technical provision and the solvency capital requirement.
We introduce the notations x, X and R for the vectors (x1, . . . , xn), (X1, . . . , Xn)
and (R1, . . . , Rn), respectively. Furthermore, the inequality ‘x > R’ is used to denote
the componentwise order. The Frechet space R (F1, . . . , Fn) is defined as the class of all
n-dimensional random vectors with fixed marginal distributions Fi, for i = 1, . . . , n. In
particular, we shall denote X ∈ R (FX1 , . . . , FXn) if Fi = FXi, where FXi
is the distribution
function of Xi, for i = 1, . . . , n.
Modern regulation and supervision should not be solely concerned with micropruden-
tial risk management, but also with macroprudential risk management. From a macro-
prudential perspective, the regulatory authority is facing, and supposed to also monitor,
the random vector
(X1 −R1, . . . , Xn −Rn).
In full generality, a risk measure is a mapping ρ from a set X of real-valued r.v.’s to
the extended real line, R:
ρ : X → R : X ∈ X 7→ ρ[X].
We denote by F−1X the left-continuous generalized inverse distribution function of X under
5
P:
F−1X (p) = infx ∈ R | FX(x) ≥ p = supx ∈ R | FX(x) < p, p ∈ [0, 1], (1)
where inf∅ = ∞ and sup∅ = −∞ by convention. In quantitative risk management,
F−1X (p) is commonly referred to as the Value-at-Risk (VaR) of X at probability level p,
denoted by VaRp[X].
Next, for a given r.v. X, we define its Tail-Value-at-Risk (TVaR) at probability level
p (sometimes referred to as Average Value-at-Risk (AVaR), Conditional Value-at-Risk
(CVaR) or Expected Shortfall (ES)), denoted by TVaRp[X], as
TVaRp[X] =1
1− p
∫ 1
p
VaRq[X] dq, p ∈ [0, 1). (2)
For further details on the properties of these two commonly used measures, their appealing
and appalling properties, and their alternatives, we refer to Denuit et al. (2005, 2006),
Dhaene et al. (2006), Goovaerts et al. (2010), Follmer and Schied (2011), and Laeven and
Stadje (2013).
2.2 Conditional stochastic ordering
In this subsection, we recall the definitions of some conditional stochastic orders, which
were first introduced by Christofides and Hadjikyriakou (2015). Up to now, these useful
notions have rarely been cited or employed in the literature. In Sections 3 and 4, we
shall employ these conditional stochastic orders and compare them with the systemic
contribution order and the systemic relevance order that we introduce.
Definition 1 Let G be a sub-σ-algebra of F . The r.v. X is said to be smaller than the
r.v. Y in the
(i) G-convex order (denoted by X ≤G−cx Y ), if E [g (X) | G] ≤ E [g (Y ) | G] holds for
every convex function g : R→ R;
(ii) G-stop-loss order (denoted by X ≤G−sl Y ), if E [g (X) | G] ≤ E [g (Y ) | G] holds for
every increasing convex function g : R→ R;
(iii) G-stochastic dominance order (denoted by X ≤G−st Y ), if E [g (X) | G] ≤ E [g (Y ) | G]
holds for every increasing function g : R→ R.
6
In case G = ∅,Ω, the conditional orders in Definition 1 reduce to the conventional
definitions of convex order, stop-loss order, and stochastic dominance order, respectively;
see the Appendix. The following implications, which are straightforward to prove, can be
found in Christofides and Hadjikyriakou (2015):
X ≤G−st [G−sl, G−cx] Y =⇒ X ≤st [sl, cx] Y. (3)
In addition, it is straightforward to verify that the following relation holds:
X ≤G−st Y =⇒ X ≤G−sl Y,
where the implication also holds if ‘≤G−st’ is replaced by ‘≤G−cx’.Hereafter, we will consider stop-loss (and other) conditional orders, based on the sub-
σ-algebra G generated by a r.v. Z. In this case, we will denote the conditional order
relation ‘≤G−sl’ by ‘≤Z−sl’.
3 Systemic contribution order
In this section, we introduce a partial order of financial institutions in terms of their
contribution to systemic risk. We consider a market composed of n (non-collaborating,
competitive) financial institutions. The potential losses over the coming reference period
(of one year, say) are denoted by X. The respective microprudential regulations are
denoted by R.
The aggregate ∆i-residual loss in the financial market is given by
SR =n∑i=1
∆i (Xi −Ri) , (4)
where ∆i : R → R, is assumed to be increasing and convex such that ∆i(0) = 0, for
i = 1, . . . , n. In the case ∆i(x) ≡ ∆(x+) with increasing convex ∆ such that ∆(0) = 0,
for i = 1, . . . , n, we shall call SR the aggregate ∆-residual loss, with its expression given
by
SR =n∑i=1
∆ ((Xi −Ri)+) . (4′)
Examples of ∆ include ∆(x) = x, and the exponential loss distance function ∆(x) =1α
[exp(αx) − 1], for α > 0. In particular, if ∆i(x) ≡ x+ for i = 1, . . . , n, SR is called as
7
the aggregate residual loss, given by
SR =n∑i=1
(Xi −Ri)+ . (4′′)
3.1 Base definition and properties
In this subsection, we introduce the systemic contribution order (base definition) by con-
ditioning on the event that the aggregate residual loss in the market exceeds a minimum
level, i.e., SR > s. We consider the special function ∆i(x) ≡ x+, for i = 1, . . . , n, which is
very meaningful in practice, and represents the net residual loss above the microprudential
risk capital. The aggregate residual loss SR in the market is then defined by (4′′).
Definition 2 (Systemic contribution order—base definition) Consider the market
of losses X, the microprudential regulation R, the aggregate residual loss SR defined in
(4′′), and the aggregate residual loss level s ∈ R+. Individual loss Xj is said to be “smaller
in systemic contribution order” than individual loss Xk under microprudential regulation
R and aggregate loss level s, denoted by Xj ≤(R,s)−con Xk, if[(Xj −Rj)+ | SR > s
]≤sl
[(Xk −Rk)+ | SR > s
].
The definition above can be interpreted as follows. The conditional individual residual
loss of institution j, given that the aggregate residual loss exceeds level s, is smaller in
terms of the stop-loss order than the corresponding conditional individual residual loss of
institution k. In that sense, institution k ‘contributes more’ to the aggregate shortfall in
the market after an aggregate residual loss of at least s.
Let us now introduce the indicator variable I (R, s), which equals 1 if the aggregate
residual loss SR exceeds s and 0 otherwise:
I (R, s) =
0, if SR ≤ s,1, if SR > s.
Based on this indicator variable, we employ the conditional stop-loss order ‘≤I(R,s)−sl’ as
defined in Subsection 2.2 to establish a relation between the systemic contribution order
and the conditional stop-loss order. The next theorem states that the conditional order
‘≤I(R,s)−sl’ between (Xj −Rj)+ and (Xk −Rk)+ is stronger than the systemic contribution
order ‘≤(R,s)−con’ between Xj and Xk given in Definition 2.
8
Theorem 3 Consider the market of losses X, the microprudential regulation R, the ag-
gregate residual loss SR defined in (4′′), and the aggregate loss level s ∈ R+. Then,
(Xj −Rj)+ ≤I(R,s)−sl (Xk −Rk)+ =⇒ Xj ≤(R,s)−con Xk.
Proof. Suppose that (Xj −Rj)+ ≤I(R,s)−sl (Xk −Rk)+. This inequality can be rewritten
in terms of the following inequality:
E[g((Xj −Rj)+
)| I (R, s)
]≤ E
[g((Xk −Rk)+
)| I (R, s)
],
which has to hold for every increasing convex function g : R → R. This implies in
particular that
E[g((Xj −Rj)+
)| I (R, s) = 1
]≤ E
[g((Xk −Rk)+
)| I (R, s) = 1
]holds for every increasing convex g. Hence, we can conclude that Xj ≤(R,s)−con Xk.
As a special case, we consider the systemic contribution order based on an event that
at least one market participant exhibits a shortfall, i.e., SR > 0.
Definition 4 (Systemic contribution order—special case) Consider the market of
losses X, the corresponding microprudential regulation R, and the aggregate residual loss
SR defined in (4′′). Individual loss Xj is said to be “smaller in systemic contribution order”
than individual loss Xk under microprudential regulation R, denoted by Xj ≤R−con Xk, if[(Xj −Rj)+ | SR > 0
]≤sl
[(Xk −Rk)+ | SR > 0
].
Similar to Definition 2, Definition 4 can be given the following interpretation. The
conditional individual residual loss of institution j, given that at least one institution in
the market is in financial distress (‘ruin’), is smaller than the corresponding conditional
individual residual loss of institution k in the sense of the stop-loss order. In that sense,
institution k ‘contributes more’ to the aggregate shortfall in the market after a collapse
of at least one of its financial entities, whence the name systemic contribution order.
The next theorem presents necessary and sufficient conditions for the systemic contri-
bution order, which strengthens the result of Theorem 3 for the special case that s = 0.
9
Theorem 5 Consider the market of losses X, the microprudential regulation R, and the
aggregate residual loss SR defined in (4′′) with P [SR > 0] > 0. Then,
Xj ≤R−con Xk ⇐⇒ (Xj −Rj)+ ≤sl (Xk −Rk)+ . (5)
Proof. The contribution order relation Xj ≤R−con Xk is equivalent to
E[(
(Xj −Rj)+ − d)+| SR > 0
]≤ E
[((Xk −Rk)+ − d
)+| SR > 0
],
which has to hold for any d. Taking into account the law of total probability and the fact
that P [SR > 0] > 0, the inequality above can be rewritten as
E[(
(Xj −Rj)+ − d)+
]≤ E
[((Xk −Rk)+ − d
)+
], for any d.
This proves the stated result.
Notice that in general, there exists no equivalence relation between Xj ≤(R,s)−con Xk
and (Xj −Rj)+ ≤sl (Xk −Rk)+, except in the case that s = 0, in which case Theorem 5
applies. This theorem states in particular that the systemic contribution order between
two market participants is then equivalent to the stop-loss order of their respective residual
losses.
Hereafter, we shall denote ‘≤I(R,s)−sl’ as ‘≤I(R)−sl’ for the special case s = 0. Next, we
prove that the conditional stop-loss order ‘≤I(R)−sl’ is equivalent to the systemic contri-
bution order ‘≤R−con’.
Theorem 6 Consider the market of losses X, the microprudential regulation R, and the
aggregate residual loss SR defined in (4′′). Then,
Xj ≤R−con Xk ⇐⇒ (Xj −Rj)+ ≤I(R)−sl (Xk −Rk)+ . (6)
Proof. If P [SR > 0] > 0, then the ‘⇐=’ implication follows immediately from (3) and
Theorem 5. If P [SR > 0] = 0, we must have (Xj −Rj)+ = 0 for all i = 1, . . . , n, and thus
the ‘⇐=’ implication holds trivially.
Let us now assume that Xj ≤R−con Xk. This inequality can be rewritten as
E[g((Xj −Rj)+
)| IR = 1
]≤ E
[g((Xk −Rk)+
)| IR = 1
],
which has to hold for every increasing convex function g. On the other hand, we have
that
E[g((Xj −Rj)+
)| IR = 0
]= g(0) = E
[g((Xk −Rk)+
)| IR = 0
]10
holds for every increasing convex function g. Hence, we can conclude that Xj ≤R−con Xk
implies that
E[g((Xj −Rj)+
)| IR]≤ E
[g((Xk −Rk)+
)| IR]
holds for every increasing convex function g. This means that also the ‘=⇒’ implication
holds.
3.2 General definition and properties
In this subsection, we generalize Definition 2 to the case of a general increasing and
convex function ∆(x+) with ∆(0) = 0. The regulator still applies the same measurement
function ∆(x+) to all institutions in the system. This means that SR takes the form (4′).
Throughout this subsection, we adopt the same notation ‘≤(R,s)−con’ and ‘≤(R)−con’ as in
the previous subsection. It should now be understood in the following sense:
Definition 7 (Systemic contribution order—general definition) Consider the mar-
ket of losses X, the microprudential regulation R, the aggregate ∆-residual loss SR defined
in (4′), and the aggregate residual loss level s ∈ R+. Individual loss Xj is said to be
“smaller in systemic contribution order” than individual loss Xk under microprudential
regulation R and aggregate loss level s, denoted by Xj ≤(R,s)−con Xk, if
[∆ ((Xj −Rj)+) | SR > s] ≤sl [∆ ((Xk −Rk)+) | SR > s] .
In accordance with the generalized Definition 7 and the fact that the conditional stop-
loss order is preserved under increasing and convex transformations, we can generalize
Theorem 3 as follows.
Theorem 8 Consider the market of losses X, the microprudential regulation R, the ag-
gregate ∆-residual loss SR defined in (4′), and the aggregate loss level s. Then,
(Xj −Rj)+ ≤I(R,s)−sl (Xk −Rk)+ =⇒ Xj ≤(R,s)−con Xk.
The following definition corresponds to the special case of Definition 7 when s = 0.
Definition 9 Consider the market of losses X, the microprudential regulation R, and
the aggregate ∆-residual loss SR defined in (4′). Individual loss Xj is said to be “smaller
in systemic contribution order” than individual loss Xk under microprudential regulation
R, denoted by Xj ≤R−con Xk, if
[∆ ((Xj −Rj)+) | SR > 0] ≤sl [∆ ((Xk −Rk)+) | SR > 0] .
11
Next, we partially generalize the results in Theorems 5 and 6 to the case of a general
measurement function by exploiting the fact that the (unconditional) stop-loss order is
(also) preserved under increasing and convex transformations.
Theorem 10 Consider the market of losses X, the microprudential regulation R, and
the aggregate ∆-residual loss SR defined in (4′) with P [SR > 0] > 0. Then,
(Xj −Rj)+ ≤sl (Xk −Rk)+ =⇒ Xj ≤R−con Xk.
Proof. In light of Theorem 4.A.8 in Shaked and Shanthikumar (2007), we know that
(Xj −Rj)+ ≤sl (Xk −Rk)+ implies ∆((Xj −Rj)+
)≤sl ∆
((Xk −Rk)+
)for increasing
and convex ∆. Then, the proof of the desired result follows from a similar argument as
the one used in the proof of Theorem 5.
Under the setting of Theorem 10, it clearly holds that Xj ≤R−con Xk is equiv-
alent to ∆((Xj −Rj)+
)≤sl ∆
((Xk −Rk)+
), which in general does not imply that
(Xj −Rj)+ ≤sl (Xk −Rk)+.
Theorem 11 Consider the market of losses X, the microprudential regulation R, and
the aggregate ∆-residual loss SR defined in (4′). Then,
(Xj −Rj)+ ≤I(R)−sl (Xk −Rk)+ =⇒ Xj ≤R−con Xk.
Proof. First, it is easy to show that the conditional stop-loss order is preserved under in-
creasing and convex transformations, which means that (Xj −Rj)+ ≤I(R)−sl (Xk −Rk)+implies ∆
((Xj −Rj)+
)≤I(R)−sl ∆
((Xk −Rk)+
). Then, the proof follows from a similar
argument as the one made in Theorem 6.
4 Systemic relevance order
In this section, we introduce the notion of systemic relevance order as a partial order of
financial institutions, characterizing one institution to be more systemically relevant than
another one. Throughout this section, we assume ∆i : R+ → R+ to be increasing and
convex, and such that ∆i(0) = 0, for i = 1, . . . , n. The aggregate ∆i-residual loss SR is
then defined as
SR =n∑i=1
∆i
((Xi −Ri)+
). (4′′′)
12
Definition 12 (Systemic relevance order) Consider the market of losses X, the mi-
croprudential regulation R, and the aggregate ∆i-residual loss SR defined in (4′′′). In-
dividual loss Xj is said to be “smaller in systemic relevance order” (or “less systemi-
cally relevant”) than individual loss Xk under microprudential regulation R, denoted by
Xj ≤R−rel Xk, if
[SR | Xj > Rj] ≤sl [SR | Xk > Rk] .
The definition above can be interpreted as follows. The conditional aggregate residual
loss in the market, given that institution j is in financial distress, is smaller (in terms of
stop-loss order), than the conditional aggregate residual loss in the market, given that
institution k is in financial distress. In that sense, institution k is ‘more relevant’ for
systemic risk.
Important to notice is that the ‘systemic relevance order’ does not order in terms of
‘contribution’ of the institution to the aggregate residual loss. A small institution could,
for example, be very ‘relevant’ in terms of systemic risk in the sense that the failure of
the institution is highly ‘connected’ to the failure of the big players in the market, but
at the same time this small institution itself will not contribute heavily to the aggregate
systemic loss. The following theoretical example illustrates this fact.
Example 13 Let (X1, X2) be a discrete random vector such that p00 = P[X1 = 0, X2 =
0] = 0.1, p01 = P[X1 = 0, X2 = 3] = 0.3, p10 = P[X1 = 1, X2 = 0] = 0.1 and p11 =
P[X1 = 1, X2 = 3] = 0.5. Suppose the microprudential risk capitals R1 and R2 for X1
and X2 are given by R1 = 0.9 and R2 = 2.8. Assume that s ∈ [0, 0.1) and ∆i(x) = x+,
for i = 1, . . . , n, which means that SR is defined by (4′′). We have that
P[(X1 −R1)+ > t|SR > s] =
p10+p111−p00 , for t ∈ [0, 0.1];
0, for t ∈ (0.1,+∞].
Similarly,
P[(X2 −R2)+ > t|SR > s] =
p01+p111−p00 , for t ∈ [0, 0.2];
0, for t ∈ (0.2,+∞].
From the expressions above, we find that
P[(X1 −R1)+ > t|SR > s] ≤ P[(X2 −R2)+ > t|SR > s], for all t ∈ R+,
which is equivalent to
[(X1 −R1)+|SR > s)] ≤st [(X2 −R2)+|SR > s)] .
13
Since stochastic dominance order implies stop-loss order, we find that X1 ≤(R,s)−con X2.
On the other hand, denoting the survival functions of [SR|X1 > R1] and [SR|X2 > R2]
by F 1 and F 2, respectively, we find that
F 1(t) =
1, for t ∈ [0, 0.1];p11
p10+p11= 5
6, for t ∈ (0.1, 0.3];
0, for t ∈ (0.3,+∞),
and
F 2(t) =
1, for t ∈ [0, 0.2];p11
p01+p11= 5
8, for t ∈ (0.2, 0.3];
0, for t ∈ (0.3,+∞).
It is straightforward to prove that∫∞tF 1(x)dx ≥
∫∞tF 2(x)dx, for any t ∈ R+, which
means that
[SR|X1 > R1] ≥sl [SR|X2 > R2] .
In other words, we have found that X2 ≤R−rel X1. We can conclude that X1 is smaller
than X2 in systemic contribution order, but that at the same time X1 is larger than X2 in
systemic relevance order. This means that X1 is more systemically relevant, but in case
of a market shortfall (SR > 0), it contributes less to the aggregate residual loss.
Let us now consider some special dependency structures among the losses in the mar-
ket, and analyze how these dependency structures play a role in relation to the systemic
relevance order. We will call the event ‘Xi > Ri’ ruin and P[Xi > Ri] the ruin probability,
where ‘ruin’ should be understood as ‘shortfall’. Hereafter, we assume that the distri-
butions of losses are continuous and strictly increasing to avoid unnecessary technical
discussions.
Theorem 14 (Comonotonic losses with identical ruin probabilities) Consider the
market of losses X, the microprudential regulation R, and the aggregate ∆i-residual loss
SR defined in (4′′′). Suppose that (Xj, Xk) =(F−1Xj
(U), F−1Xk(U))
with U a standard uni-
form r.v., and FXj(Rj) = FXk
(Rk), for some 1 ≤ j 6= k ≤ n. Then,
[SR | Xj > Rj] = [SR | Xk > Rk] .
Proof. The event that institution j ruins, i.e., Xj > Rj, can be equivalently expressed
as U > FXj(Rj). Indeed,
Xj > Rj ⇐⇒ F−1Xj(U) > Rj ⇐⇒ U > FXj
(Rj).
14
Similarly, we have that Xk > Rk ⇐⇒ U > FXk(Rk). Hence, the ruin of institution j is
equivalent to the ruin of institution k due to FXj(Rj) = FXk
(Rk).
The theorem above states that in a market with microprudential risk capitals having
the same ruin probabilities, any two institutions that are comonotonic are equally rele-
vant in terms of systemic risk. In particular, we have that in a VaR-based comonotonic
market, all institutions are equally relevant in terms of systemic risk provided that all
microprudential risk capitals have the same confidence levels.
Next, we revisit Theorem 14 when the confidence levels for microprudential regulations
are different. In the following theorem, we will use the hazard rate order, notation ‘≤hr’,which is defined in the Appendix.
Theorem 15 (Comonotonic losses with different ruin probabilities) Consider the
market of losses X, the microprudential regulation R, and the aggregate ∆i-residual loss
SR defined in (4′′′). Suppose that (X1, . . . , Xn) =(F−1X1
(U), . . . , F−1Xn(U))
with U standard
uniform. If FXj(Rj) ≥ FXk
(Rk) for some 1 ≤ j 6= k ≤ n, then
[SR | Xj > Rj] ≤hr [SR | Xk > Rk] .
Proof. Note that Xj > Rj = F−1Xj(U) > Rj = U > FXj
(Rj) and Xk > Rk =
U > FXk(Rk). Thus, it follows that Xj > Rj ⊆ Xk > Rk. According to Theorem
1.B.20 of Shaked and Shanthikumar (2007), it holds that
[U | Xj > Rj] = [U | U > FXj(Rj)] ≤hr [U | U > FXk
(Rk)] = [U | Xk > Rk].
Since SR is an increasing function of U , from Theorem 1.B.2 of Shaked and Shanthikumar
(2007) we have
[SR | Xj > Rj] ≤hr [SR | Xk > Rk] ,
which yields the desired result.
It should be mentioned that the result of Theorem 15 can be strengthened to the
likelihood ratio order (Section 1.C in Shaked and Shanthikumar, 2007) by using Theorems
1.C.8 and 1.C.27 in Shaked and Shanthikumar (2007). Since the hazard rate order implies
the stop-loss order, the above theorem implies that in a VaR-based comonotonic market
with microprudential risk capitals having different confidence levels, the institution with
the lower confidence level is more systemically relevant than the institution with the larger
confidence level.
15
Next, we consider the situation where the market consists of two identically distributed
counter-monotonic losses having a common microprudential regulation, and where the
regulator uses the same measurement function for each loss.
Theorem 16 (Counter-monotonic losses with same ruin probability) Consider a
market of two losses (X1, X2) = (F−1(U), F−1(1− U)) for some cdf F with U a standard
uniform r.v., with microprudential regulation R = (R,R) where 12< F (R) < 1, and with
the aggregate ∆-residual loss SR defined in (4′). Then,
[SR | X1 > R]d= [SR | X2 > R] .
Proof. The event that institution 1 ruins, i.e., X1 > R, can be equivalently expressed as
U > F (R). Indeed,
X1 > R⇐⇒ F−1(U) > R⇐⇒ U > F (R).
Similarly, the event that institution 2 fails, i.e., X2 > R, can be equivalently expressed as
U < 1− F (R). Since 12< F (R) < 1, we have 1− F (R) < F (R). Hence, we find that
[SR | X1 > R] = [SR | U > F (R)]
= [∆ (X1 −R) | U > F (R)]
=[∆(F−1(U)−R
)| U > F (R)
].
On the other hand, we have that
[SR | X2 > R] = [SR | U < 1− F (R)]
= [∆ (X2 −R) | U < 1− F (R)]
=[∆(F−1(1− U)−R
)| 1− U > F (R)
].
This proves the stated result.
The theorem above states that in a market consisting of two institutions with counter-
monotonic but identically distributed risks and identical microprudential regulation, each
institution is equally relevant in terms of systemic risk.
Let us now consider a market consisting of two institutions with counter-monotonic and
identically distributed stochastic losses, but with different microprudential risk capitals.
Recall that a r.v. X or its distribution F is said to be IFR [DFR] if − logF (x) is convex
[concave] (see p. 31 in Denuit et al., 2005).
16
Theorem 17 (Counter-monotonic losses with different ruin probabilities) Consider
a market of two losses (X1, X2) = (F−1(U), F−1(1− U)) for some cdf F with U a standard
uniform r.v., with microprudential regulation R = (R1, R2) where 0 < F (R1), F (R2) < 1
such that F (R1) + F (R2) > 1 and R1 > R2. Furthermore, consider the aggregate ∆-
residual loss SR defined in (4′). If F is IFR [DFR], then
[SR | X1 > R1] ≤hr [≥hr] [SR | X2 > R2] .
Proof. Note that X1 > R1 ⇐⇒ U > F (R1) and X2 > R2 ⇐⇒ U < 1 − F (R2). Then,
from F (R1) + F (R2) > 1, it follows that
[SR | X1 > R1] = [SR | U > F (R1)]
= [∆ (X1 −R1) | U > F (R1)]
=[∆(F−1(U)−R1
)| F−1(U) > R1
]and
[SR | X2 > R2] = [SR | U < 1− F (R2)]
= [∆ (X2 −R2) | U < 1− F (R2)]
=[∆(F−1(1− U)−R2
)| 1− U > F (R2)
]d=
[∆(F−1(U)−R2
)| U > F (R2)
]=
[∆(F−1(U)−R2
)| F−1(U) > R2
].
In light of R1 > R2, the desired result follows from Theorem 1.B.38 and Theorem 1.B.2
in Shaked and Shanthikumar (2007).
If F has a log-concave density, the result ‘≤hr’ in Theorem 17 can be strengthened to
the likelihood ratio order by using Theorem 1.C.52 of Shaked and Shanthikumar (2007).
The theorem above states that in a market of two institutions with counter-monotonic
identically IFR-distributed risks with different risk capitals, the institution with the larger
microprudential risk capital is less systemically relevant than the other institution. The
opposite conclusion holds for DFR-distributed risks.
To conclude this section, we present sufficient conditions for independent individual
financial institutions to be ordered in the systemic relevance order sense.
17
Theorem 18 (Independent losses with identical ruin probabilities) Consider a mar-
ket of mutually independent losses X, microprudential regulation R, and aggregate ∆-
residual loss SR as defined in (4′). If FXj(Rj) = FXk
(Rk) for some 1 ≤ j 6= k ≤ n,
then
(Xj −Rj)+ ≤sl (Xk −Rk)+ =⇒ Xj ≤R−rel Xk. (7)
Proof. We introduce the notation
SRj ,Rk= ∆
((Xj −Rj)+
)+ ∆
((Xk −Rk)+
).
From the stop-loss ordering relation (7) and the increasingness and convexity of ∆, we
find that
∆((Xj −Rj)+
)≤sl ∆
((Xk −Rk)+
).
This observation implies that(1− FXj
(Rj))E[(SRj ,Rk
− d)+| Xj > Rj
]= E
[(SRj ,Rk
− d)+
]− FXj
(Rj)E[(SRj ,Rk
− d)+| Xj ≤ Rj
]= E
[(SRj ,Rk
− d)+
]− FXj
(Rj)E[(
∆((Xk −Rk)+
)− d)+
]≤ E
[(SRj ,Rk
− d)+
]− FXk
(Rk)E[(
∆((Xj −Rj)+
)− d)+
]= (1− FXk
(Rk))E[(SRj ,Rk
− d)+| Xk > Rk
].
Hence, we have proven that[SRj ,Rk
| Xj > Rj
]≤sl
[SRj ,Rk
| Xk > Rk
]. Taking into ac-
count the mutual independence between the components of X, we find that
[SR | Xj > Rj] =
n∑i=1i 6=j,k
∆((Xi −Ri)+
)+ SRj ,Rk
| Xj > Rj
≤sl
n∑i=1i 6=j,k
∆((Xi −Ri)+
)+ SRj ,Rk
| Xk > Rk
= [SR | Xk > Rk] ,
which proves that Xj ≤R−rel Xk.
The theorem states the intuitive result that in a market of players with mutually
independent losses and with microprudential VaR-based risk capitals having the same
18
confidence levels for any two institutions j and k, the institution with the largest individual
residual risk (in terms of stop-loss order) is more systemically relevant.
Remark 19 In Theorems 14-18, we derived sufficient conditions for the systemic rel-
evance order based on the idea to restate the aggregate market event ‘SR > s’ in a
simpler form, making use of the institution-specific conditioning events ‘Xj > Rj’, for
j = 1, . . . , n. Notice that such a procedure is in general not possible for the systemic
contribution order, as the individual r.v.’s (Xj − Rj)+ cannot be restated based on the
market-based conditioning event ‘SR > s’. For both orders—systemic contribution and
systemic relevance order—there are a r.v. as well as a conditioning event on both the
LHS and the RHS of the inequality. For the systemic relevance order, the r.v.’s are equal
on the LHS and the RHS, but the conditioning event is different. For the systemic contri-
bution order, the r.v.’s are different on the LHS and the RHS, but the conditioning event
is the same. Now, in the comonotonic case, one can easily show that for the systemic
relevance order, the two conditioning events become explicitly linked or even identical,
implying that we have a related or even the same conditional object on both the LHS and
the RHS, while this is not the case for the systemic contribution order.
5 Systemic risk order
In this section, we define and investigate the systemic risk order. This stochastic order
allows us to compare financial systems which only differ in terms of the dependence
structure between the residual losses. As before, for any i ∈ 1, . . . , n, we assume the
measurement function ∆i to be increasing and convex unless otherwise mentioned.
5.1 Definition
Definition 20 (Systemic risk order) Consider X and Y , both elements of the Frechet
space R (F1, . . . , Fn). Furthermore, consider the microprudential regulation given by R.
Then, X is said to be smaller than Y in “systemic risk order” under microprudential
regulation R, denoted by X ≤R−sr Y ), if the aggregate ∆i-residual loss of X is smaller in
stop-loss order than the aggregate ∆i-residual loss of Y :
X ≤R−sr Y ⇐⇒n∑i=1
∆i (Xi −Ri) ≤sln∑i=1
∆i (Yi −Ri) . (8)
19
Furthermore, we say that X is smaller than Y in systemic risk order, denoted by X ≤sr Y ,
if X ≤R−sr Y for any R.
Remark 21 The systemic contribution order and the systemic relevance order, which
were introduced in Sections 3 and 4, respectively, partially order financial institutions in
a given financial system (i.e., a system of n institutions with given individual loss dis-
tributions and given dependence structure connecting these individual loss distributions).
On the other hand, the systemic risk order considers the Frechet space of all financial
systems with given individual loss distributions. This order allows us to partially order
financial systems in the given Frechet space based on the dependency structure among the
individual losses.
Under the setup of Definition 20, we know that the microprudential regulation R is
fixed and X and Y have the same marginals for the systemic risk order ‘≤R−sr’. Hence,
X ≤R−sr Y means that Y is more positively interconnected and thus leads to larger
aggregate ∆i-residual loss in the stop-loss order sense. The absence of a macroprudential
regulation implies that the markets X and Y , with X ≤R−sr Y , are treated equally,
while obviously, the second situation is the more dangerous one. In order to overcome
this inconsistency, one should set up a regulation with a microprudential as well as a
macroprudential leg. How to set up a macroprudential regulation is the research topic in
our forthcoming companion paper (Dhaene et al., 2019b).
The definition of systemic risk order can be extended to compare random vectors of a
broader class in case the micro-level capital requirements Ri’s only depend on an upper
tail of the distribution. Suppose that Ri only depends onF−1i (q) | q ≥ p
. Then we can
define the ‘≤R−sr’ order between members of the class of all n-dimensional distributions
with fixed tailsF i(xi) | xi > F−1i (p)
, i = 1, 2, . . . , n.
5.2 Some basic properties of the systemic risk order
In this subsection, we present some sufficient conditions imposed on the random vec-
tors X and Y in the same Frechet space to be ordered in the systemic risk order.
Note that X ≤sr Y , that is, X ≤R−sr Y for any R, is equivalent to requiring that∑ni=1 fi(Xi) ≤sl
∑ni=1 fi(Yi), for any increasing and convex fi, which is a natural general-
ization (strengthening) of∑n
i=1Xi ≤sl∑n
i=1 Yi.
Theorem 22 Assume that X, Y ∈ R(F1, . . . , Fn). Then, we have that:
20
(i) Systemic risk is highest when the losses in the market are comonotonic:
X ≤sr(F−11 (U) , . . . , F−1n (U)
).
(ii) Supermodular order implies systemic risk order:
X ≤sm Y =⇒ X ≤sr Y .
(iii) Multivariate stop-loss order implies systemic risk order:
X ≤sl Y =⇒ X ≤sr Y .
(iv) Finally, a more severe micro-level regulation leads to less aggregate systemic risk:
R ≤ R′ =⇒n∑i=1
∆i
(Xi −R
′
i
)≤
n∑i=1
∆i (Xi −Ri) .
Proof. For (i), it is easy to see that (F−11 (U), . . . , F−1n (U)) is a comonotonic vector
contained in R(F1, . . . , Fn), which implies that (∆1(F−11 (U)−R1), . . . ,∆n(F−1n (U)−Rn))
is the comonotonic counterpart of (∆1(X1 − R1), . . . ,∆n(Xn − Rn)). Hence, the desired
result follows from the fact that the sum of a vector in Frechet space is maximized (in
terms of the convex order) in the comonotonic case; see Corollary 3.4.30 in Denuit et al.
(2005).
By using Definition 8, the proofs of (ii) and (iii) follow from Propositions 6.3.9 and
3.4.65 in Denuit et al. (2005), respectively.
The proof of (iv) is easily obtained from the increasingness of ∆i and the condition
that R′i ≤ Ri, for i = 1, . . . , n.
The systemic risk order can be identified as a specific version of the multivariate stop-
loss order in which the increasing convex functions g : Rn → R are restricted to the class
of linear-convex functions. Interestingly, Koshevoy and Mosler (1996, 1997, 1998) restrict
to the complementary case of convex-linear functions in a similar setting.
Recall that X is said to be R-upper comonotonic if ((X1 − R1)+, . . . , (Xn − Rn)+)
is comonotonic. This concept and related properties were introduced and discussed in
Cheung (2009), Dong et al. (2010) and Nam et al. (2011). Let us for the moment suppose
that Ri = F−1i (0.995). Then, the most dangerous situation (in terms of systemic risk)
occurs when X is upper comonotonic at level 0.995. In this case, a shortfall of one
21
of the institutions in the sense that Xi > F−1i (0.995), is accompanied by a shortfall
of all institutions. Again, in existing regulation, where only a microprudential capital
requirement applies, no distinction is made between this “explosive” situation and the
less frightening situation where all tails (above F−1i (0.995)) are independent.
For special distance measures, weaker requirements already yield special partial or-
dering results. This is shown in the following proposition.
Proposition 23 Suppose ∆i(x) = ∆i(x+), for i = 1, . . . , n, and X, Y ∈ R(F1, . . . , Fn).
Then, we have:
(i) Systemic risk is highest when the losses in the market are upper-comonotonic:
n∑i=1
∆i
((Xi −Ri)+
)≤sl
n∑i=1
∆i
((Xi −Ri)
c+
).
(ii) “Upper” supermodular order implies systemic risk order:
(X −R)+ ≤sm (Y −R)+ =⇒ X ≤R−sr Y .
(iii) “Upper” multivariate stop-loss order implies systemic risk order:
(X −R)+ ≤sl (Y −R)+ =⇒ X ≤R−sr Y .
Proof. (i) holds as a consequence of Theorem 1 in Dong et al. (2010). (ii) and (iii) follow
from Propositions 6.3.9 and 3.4.65 in Denuit et al. (2005), respectively.
5.3 Majorization, diversity and systemic risk
The notion of majorization, characterizing the diversity of coordinates of real vectors, is
useful in establishing various inequalities arising from actuarial science, applied probability
as well as reliability theory. Let x1:n ≤ · · · ≤ xn:n be the increasing arrangement of the
components of the vector x = (x1, . . . , xn).
Definition 24 A vector x ∈ Rn is said to majorize another vector y ∈ Rn, written as
xm
y, if∑j
i=1 xi:n ≤∑j
i=1 yi:n for j = 1, . . . , n− 1, and∑n
i=1 xi:n =∑n
i=1 yi:n.
Pan et al. (2015) studied stochastic properties of∑n
i=1 φ(Xi, ai) in the sense of the
stochastic dominance order and the stop-loss order, under some additional conditions
22
when the joint density fX(x) of the r.v.’s Xi is arrangement increasing (AI), where φ is
a bivariate function and ai is an indexing parameter of the r.v. Xi, for i = 1, . . . , n. The
notion of stochastic arrangement increasing (SAI) is introduced in Cai and Wei (2014)
and depicts a positive dependence structure of the components of a random vector. For
an absolutely continuous random vector, SAI is equivalent to the statement that the joint
density function is an AI function. Many multivariate distributions have an AI density,
including the multivariate versions of the Dirichlet distribution, the inverted Dirichlet
distribution, the F distribution, and the Pareto distribution of type I (see Hollander
et al., 1977). In particular, Pan et al. (2015) showed that am
b implies∑n
i=1 φ(Xi −a(n−i+1):n) ≥st
∑ni=1 φ(Xi − b(n−i+1):n) provided that fX(x) is log-concave, arrangement
increasing and φ is a convex function. If fX(x) is arrangement increasing and φ is convex,
we have that am
b implies∑n
i=1 φ(Xi − a(n−i+1):n) ≥sl∑n
i=1 φ(Xi − b(n−i+1):n).
Right tail weakly stochastic arrangement increasing (RWSAI), which is weaker than
SAI, is a nonparametric positive dependence notion, introduced by Cai and Wei (2014),
depicting that the r.v.’s in the random vector are not only positively dependent, but also
ordered in some stochastic sense. For further discussions and applications of this concept
and other related dependence notions, we refer to Cai and Wei (2014, 2015) and Zhang
et al. (2018). Recently, the above result of Pan et al. (2015) has been generalized to the
case of RWSAI X by Proposition 5.1 of You and Li (2015).
Consider two different configurations of the microprudential regulationR = (R1, . . . , Rn)
and R′ = (R′1, . . . , R′n). Then, by applying Corollary 3.8, Theorem 3.12(ii) in Pan et al.
(2015) and Proposition 5.1 of You and Li (2015), the following result is obtained imme-
diately if ∆i ≡ ∆ is increasing and convex, for i = 1, . . . , n.
Theorem 25 Let R = (R1, . . . , Rn) and R′ = (R′1, . . . , R′n), while SR =
∑ni=1 ∆(Xi−Ri)
and SR′ =∑n
i=1 ∆(Xi −R′i). Assume that R1 ≥ R2 ≥ . . . ≥ Rn.
(i) If fX(x) is log-concave and AI, then Rm
R′ implies SR ≥st SR′.
(ii) If X is RWSAI, then Rm
R′ implies SR ≥sl SR′.
Proof. Since ∆ is increasing and convex, we have that ∆(x− y) is submodular in (x, y).
By using a similar proof as the one of Proposition 3.7 in Pan et al. (2015), while taking
into account the results in Theorem 3.12(ii) in Pan et al. (2015) and Proposition 5.1 of
You and Li (2015), the desired result is obtained immediately.
23
The conditions (i) and (ii) in the theorem above mean that the n losses are arrayed
in ascending order. If a larger loss is accompanied by a smaller risk capital, more hetero-
geneity among the risk capitals leads to greater aggregate ∆-residual loss in the sense of
the stochastic dominance order or the stop-loss order.
5.4 Aggregate residual loss and conditioning events
Finally, we discuss the stochastic properties of the aggregate ∆i-residual loss when condi-
tioned upon some systemic risk events. It is common practice to evaluate risks condition-
ally upon stress scenarios; see also Dhaene et al. (2019a,b). Furthermore, the choice of the
risk measure used to evaluate SR may include a choice of a conditioning event. From this
perspective it is relevant to investigate the behavior of SR with respect to conditioning
events.
Theorem 26 The following statements hold:
(i) X is associated ⇐⇒ ∆(X −R) = (∆1(X1 − R1), . . . ,∆n(Xn − Rn)) is associated
for all ∆.
Let X be associated. Then:
(ii) SR ≤st [SR|A], for aggregate ∆i-residual loss SR defined in (4) and systemic risk
event A = ∆i (Xi −Ri) > t, for some t ≥ 0 and i = 1, 2, . . . , n.
(iii) [SR|A1] ≤st [SR|A2], for aggregate ∆i-residual loss SR defined in (4) and systemic
risk events
Aj = ∆i (Xi −Ri) > tj, for some tj ≥ 0 and i = 1, 2, . . . , n,
for j = 1, 2 and t1 < t2.
Proof. Proof of (i): If X is associated, then ∆(X −R) is also associated since asso-
ciatedness is preserved under component-wise increasing transformations. The converse
follows trivially.
Proof of (ii): Since ∆(X −R) is associated, it follows that ∆(X −R) ≤st [∆(X −R)|A]
for the risk event A = ∆i (Xi −Ri) > t, for i = 1, 2, . . . , n by using Theorem 3.1 of
Colangelo et al. (2008). Then, the statement is proved by applying Theorem 6.B.16 of
Shaked and Shanthikumar (2007).
24
Proof of (iii): By treating the conditional random vector [∆(X −R)|A1] as a new vec-
tor, the proof can be completed by using a similar argument as in (ii) since [∆(X −R)|A2] =
[[∆(X −R)|A1]|A2].
If we define the conditioning event as “the aggregate ∆i-residual loss SR exceeds a
certain threshold”, we find the following result.
Theorem 27 Let B = SR > t, for some t ≥ 0. The following statements hold:
(i) SR ≤hr [SR|B], for aggregate ∆i-residual loss SR defined in (4).
(ii) [SR|B1] ≤hr [SR|B2], for aggregate ∆i-residual loss SR defined in (4) and systemic
risk events Bj = SR > tj, for some tj ≥ 0, for j = 1, 2 and t1 < t2.
Proof. The proof of (i) follows directly from Theorem 1.B.20 of Shaked and Shanthikumar
(2007). For (ii), the proof can be obtained by using a similar argument as in Theorem
26(iii) and the result in Theorem 1.B.20 of Shaked and Shanthikumar (2007).
The two theorems above make explicit what additional requirements need to be sat-
isfied in order for systemic risk events to increase the aggregate ∆i-residual loss in the
sense of the stochastic dominance order and the hazard rate order, respectively.
6 Conclusion
In this paper, we have introduced the systemic contribution order and the systemic rel-
evance order, which are useful for investigating systemic risk in a market consisting of
several financial institutions. The systemic contribution order ‘Xj ≤(R,s)−con Xk’ indicates
that institution j contributes less systemically than institution k, in case the aggregate
∆i-residual loss exceeds level s. In particular, the systemic contribution order ‘≤R−con’
(corresponding to the case s = 0) can be characterized in terms of a conditional stop-loss
order of individual residual risks, as introduced in Christofides and Hadjikyriakou (2015).
On the other hand, the systemic relevance order is an ‘order of information’. Indeed,
Xj ≤R−rel Xk means that the information that institution k is in financial distress is more
relevant in terms of the aggregate residual loss than the corresponding information about
institution j.
We also introduced the systemic risk order which can be used to study the effect that
the interconnectedness of individual losses has on the aggregate residual risk in terms
25
of the stop-loss order. The systemic risk order compares the aggregate systemic risk
and involves the dependence structure (copula) for a given Frechet space, but not the
distributions of the individual losses.
The three new stochastic orders that we introduced in this paper may be used to
investigate and compare systemic risk in financial institutions and systems. All three
stochastic orders invoke the time-honored actuarial stop-loss order but can be straight-
forwardly adapted to other stochastic orders.
As observed from Example 13, a small institution (A) contributing less in terms of
the systemic contribution order than another one (B) may be very relevant to the whole
system in the sense of the systemic relevance order. Indeed, this small institution A
might be, e.g., highly related to a third big institution, which contributes more than
B to the systemic risk in the market. A promising research direction is how to set up
microprudential and macroprudential regulations for all individual institutions.
In a first companion paper (Dhaene et al., 2019a), we propose new conditional risk
measures to quantify conditional systemic risk events related to the systemic contribution
order and the systemic relevance order. In a second companion paper (Dhaene et al.,
2019b), we propose and investigate the properties of procedures for attributing and allo-
cating the aggregate systemic risk and organizing a scheme to assess a macroprudential
surcharge for each individual financial institution, which is fair and consistent with the
allocated risk.
Acknowledgements
We are grateful for comments from Emiliano Valdez on an early draft of this paper.
Jan Dhaene acknowledges the financial support of the Research Foundation Flanders
(FWO) under grant GOC3817N. Roger Laeven acknowledges the financial support of the
Netherlands Organization for Scientific Research under grant NWO VIDI. Yiying Zhang
acknowledges the start-up grant from Nankai University, and the financial support and
nice working place from the actuarial research group at KU Leuven and the Amsterdam
Center of Excellence in Risk and Macro Finance at the University of Amsterdam during
his visit in the fall of 2018.
References
Basel Committee on Banking Supervision. International convergence of capital measurement
26
and capital standards. Working Paper, Bank for International Settlements, 1988. Availableat http://www.bis.org/publ/bcbs04a.pdf.
Basel Committee on Banking Supervision. International convergence of capital measurementand capital standards: a revised framework comprehensive version. Working Paper, Bank forInternational Settlements, 2006. Available at http://www.bis.org/publ/bcbs128.pdf.
Basel Committee on Banking Supervision. Global systemically important banks: Assessmentmethodology and the additional loss absorbency requirements. Working Paper, Bank forInternational Settlements, 2011. Available at http://www.bis.org/publ/bcbs201.pdf.
Jun Cai and Wei Wei. Some new notions of dependence with applications in optimal allocationproblems. Insurance: Mathematics and Economics, 55:200–209, 2014.
Jun Cai and Wei Wei. Notions of multivariate dependence and their applications in optimalportfolio selections with dependent risks. Journal of Multivariate Analysis, 138:156–169, 2015.
Ka Chun Cheung. Upper comonotonicity. Insurance: Mathematics and Economics, 45(1):35–40,2009.
Tasos C Christofides and Milto Hadjikyriakou. The conditional convex order and a comparisoninequality. Stochastic Analysis and Applications, 33(2):259–270, 2015.
Antonio Colangelo, Taizhong Hu, and Moshe Shaked. Conditional orderings and positive de-pendence. Journal of Multivariate Analysis, 99:358–371, 2008.
J. David Cummins and Mary A. Weiss. Systemic risk and the insurance industry. In GeorgesDionne, editor, Handbook of Insurance, pages 745–793. Springer: New York, 2nd edition, 2013.
Michel Denuit, Jan Dhaene, Marc J. Goovaerts, and Rob Kaas. Actuarial Theory for DependentRisks. New York: Wiley, 2005.
Michel Denuit, Jan Dhaene, Marc J. Goovaerts, Rob Kaas, and Roger J. A. Laeven. Riskmeasurement with equivalent utility principles. Statistics and Decisions, Special Issue onRisk Measures: General Aspects and Applications, 24(1):1–26, 2006.
Jan Dhaene and Marc J Goovaerts. Dependency of risks and stop-loss order. ASTIN Bulletin:The Journal of the IAA, 26(2):201–212, 1996.
Jan Dhaene, Michel Denuit, Marc J. Goovaerts, Rob Kaas, and David Vyncke. The conceptof comonotonicity in actuarial science and finance: Theory. Insurance: Mathematics andEconomics, 31(1):3–33, 2002a.
Jan Dhaene, Michel Denuit, Marc J Goovaerts, Rob Kaas, and David Vyncke. The concept ofcomonotonicity in actuarial science and finance: Applications. Insurance: Mathematics andEconomics, 31(2):133–161, 2002b.
27
Jan Dhaene, Steven Vanduffel, Qihe Tang, Marc J. Goovaerts, Rob Kaas, and David Vyncke.Risk measures and comonotonicity: A review. Stochastic Models, 22(4):573–606, 2006.
Jan Dhaene, Roger J. A. Laeven, and Yiying Zhang. Systemic risk: Conditional distortion riskmeasures. Preprint, 2019a.
Jan Dhaene, Roger J. A. Laeven, and Yiying Zhang. Systemic risk: Allocation principles.Working Paper, 2019b.
Jing Dong, Ka Chun Cheung, and Hailiang Yang. Upper comonotonicity and convex upperbounds for sums of random variables. Insurance: Mathematics and Economics, 47(2):159–166, 2010.
Paul Embrechts, Andrea Hoing, and Giovanni Puccetti. Worst VaR scenarios. Insurance: Math-ematics and Economics, 37(1):115–134, 2005.
J.D. Esary, F. Proschan, and D.W. Walkup. Association of random variables, with applications.Annals of Mathematical Statistics, 38(5):1466–1474, 1967.
Hans Follmer and Alexander Schied. Stochastic Finance, 3rd ed. Berlin: De Gruyter, 2011.
Marc J. Goovaerts, Rob Kaas, and Roger J. A. Laeven. Decision principles derived from riskmeasures. Insurance: Mathematics and Economics, 47(2):294–302, 2010.
Marc J. Goovaerts, Rob Kaas, and Roger J. A. Laeven. Worst case risk measurement: Back tothe future? Insurance: Mathematics and Economics, 49(3):380–392, 2011.
Group of Ten. Report on Consolidation in the Financial Sector. 2001.
Myles Hollander, Frank Proschan, and Jayaram Sethuraman. Functions decreasing in transpo-sition and their applications in ranking problems. The Annals of Statistics, pages 722–733,1977.
Harry Joe. Multivariate Models and Multivariate Dependence Concepts. Chapman andHall/CRC, 1997.
Rob Kaas, Roger J. A. Laeven, and Roger B. Nelsen. Worst VaR scenarios with given marginalsand measures of association. Insurance: Mathematics and Economics, 44(2):146–158, 2009.
George G. Kaufman and Kenneth E. Scott. What is systemic risk, and do bank regulators retardor contribute to it? Independent Review, 7:371–391, 2003.
G. A. Koshevoy and K. Mosler. The Lorenz zonoid of a multivariate distribution. Journal ofthe American Statistical Association, 91(434):873–882, 1996.
G. A. Koshevoy and K. Mosler. Multivariate Gini indices. Journal of Multivariate Analysis, 60(2):252–276, 1997.
28
G. A. Koshevoy and K. Mosler. Lift zonoids, random convex hulls and the variability of randomvectors. Bernoulli, 4(3):377–399, 1998.
Roger J. A. Laeven. Worst VaR scenarios: A remark. Insurance: Mathematics and Economics,44(2):159–163, 2009.
Roger J. A. Laeven and Mitja Stadje. Entropy coherent and entropy convex measures of risk.Mathematics of Operations Research, 38(2):265–293, 2013.
Bo Henry Lindqvist. Association of probability measures on partially ordered spaces. Journalof Multivariate Analysis, 26(2):111–132, 1988.
Alfred Muller and Marco Scarsini. Stochastic comparison of random vectors with a commoncopula. Mathematics of Operations Research, 26(4):723–740, 2001.
Hee Seok Nam, Qihe Tang, and Fan Yang. Characterization of upper comonotonicity via tailconvex order. Insurance: Mathematics and Economics, 48(3):368–373, 2011.
Roger B. Nelsen. An Introduction to Copulas. Springer Science & Business Media, 2007.
Xiaoqing Pan, Min Yuan, and Subhash C Kochar. Stochastic comparisons of weighted sums ofarrangement increasing random variables. Statistics & Probability Letters, 102:42–50, 2015.
Moshe Shaked and J George Shanthikumar. Stochastic orders. Springer Science & BusinessMedia, 2007.
Yinping You and Xiaohu Li. Functional characterizations of bivariate weak SAI with an appli-cation. Insurance: Mathematics and Economics, 64:225–231, 2015.
Yiying Zhang, Xiaohu Li, and Ka Chun Cheung. On heterogeneity in the individual model withboth dependent claim occurrences and severities. ASTIN Bulletin: The Journal of the IAA,48(2):817–839, 2018.
29
A Appendix: Related definitions and useful proper-
ties
A.1 Univariate stochastic ordering
In this section, we recall the definitions of some univariate stochastic orders used in the
main text.
Definition 28 Let FX(x) = 1− FX(x) and F Y (y) = 1− FY (y) be the survival functions
of X and Y , respectively, and let hX and hY be their hazard rates (i.e., the ratios of the
probability density functions to the survival functions). Then, X is said to be smaller than
Y in the
(i) hazard rate order, denoted by X ≤hr Y , if F Y (t)/FX(t) is increasing in t ∈ R, or
equivalently, hY (t) ≤ hX(t) for all t ∈ R, where hX and hY are the hazard rate
functions of X and Y , respectively;
(ii) stochastic dominance order, denoted by X ≤st Y , if E[φ(X)] ≤ E[φ(Y )] for any
increasing φ : R 7→ R;
(iii) stop-loss order [or increasing convex order], denoted by X ≤sl Y , if E[(X − d)+] ≤E[(Y − d)+] for all d ∈ R+, or equivalently, E[φ(X)] ≤ E[φ(Y )] for all increasing
convex function φ : R→ R;
(iv) convex order, denoted by X ≤cx Y , if E[X] = E[Y ] and X ≤sl Y .
It is known that the hazard rate order implies the stochastic dominance order, which in
turn implies the stop-loss order. For further details on their properties and applications,
we refer to Denuit et al. (2005) and Shaked and Shanthikumar (2007).
A.2 Multivariate stochastic ordering
A.2.1 Measuring dependence
In this section, we recall the definitions of some positive dependence notions used in the
main text and indicate their connection to other positive dependence notions. A subset
A ⊆ Rn is said to be comonotonic if, for any x ∈ A and y ∈ A, either xi ≤ yi for
i = 1, . . . , n or xi ≥ yi for i = 1, . . . , n. A random vector X is said to be comonotonic if
there is a comonotonic subset A such that P[X ∈ A] = 1.
30
Consider n r.v.’s X1, . . . , Xn. Define S =∑n
i=1Xi and let Sc =∑n
i=1 F−1i (U) be the
comonotonic sum, where Fi is the distribution of Xi, for i = 1, . . . , n, and U is uniform
on (0, 1). It is known that the comonotonic random vector (F−11 (U), . . . , F−1n (U)) is
maximal within the corresponding Frechet space in the sense of the convex order of the
sum, i.e., S ≤sl Sc. This useful concept has been widely used in actuarial science to model
the strongest positive dependence structure among risks. For comprehensive discussions
on comonotonicity and its applications in insurance and finance, readers are referred to
Dhaene et al. (2002a,b).
A random vector X is said to be positively lower and upper orthant dependent (PLOD
and PUOD) if
P [X ≤ x] ≥n∏i=1
P [Xi ≤ xi] and P [X > x] ≥n∏i=1
P [Xi > xi] , ∀x ∈ Rn, (9)
respectively. The vector X is positively orthant dependent (POD) if both inequalities
in (9) hold. In the bivariate case, the two inequalities in (9) are equivalent and POD
reduces to positive quadrant dependence (PQD). Both PUOD and PLOD (hence, POD)
are preserved under component-wise increasing and continuous transformations.
It was shown in Proposition 5.3.9 of Denuit et al. (2005) that X⊥1 +X⊥2 ≤sl X1 +X2,
where (X⊥1 , X⊥2 ) is an independent version of PQD (X1, X2) having the same marginal
distributions. This result, however, does not extend to the multivariate case (n > 2) of
POD.
Next, we say that X is associated (Esary et al. (1967)) if
Cov [g(X), h(X)] ≥ 0, (10)
for all increasing functions g, h : Rn → R such that the covariance exists. Associatedness
is preserved under component-wise increasing transformations. Association implies POD
(Denuit et al. (2005), p. 319). Lindqvist (1988) provided an equivalent representation of
associatedness: X is associated if
P [X ∈ U ∩ V ] ≥ P [X ∈ U ]P [X ∈ V ] ,
for all upper sets U, V ⊆ Rn.
Furthermore, we define the notions of conditional increasingness (CI) and conditional
increasingness in sequence (CIS); see Muller and Scarsini (2001). We say that X is CIS
if, for all i = 2, 3, . . . , n,
Xi|X1 = x1, . . . , Xi−1 = xi−1 ≤st Xi|X1 = y1, . . . , Xi−1 = yi−1,
31
whenever yj ≥ xj, assumed to be in the support of Xj, j = 1, . . . , i − 1. Based on this
notion, we say that X is CI if (Xπ(1), . . . , Xπ(n)) is CIS for all permutations π of 1, . . . , n.Of course, CI implies CIS.
Finally, we give the notion of multivariate total positivity of order 2 (MTP2). Suppose
X has a continuous or discrete density fX . Then, X is MTP2 if log fX is supermodular.
(A function g : Rn → R is supermodular if g(x) + g(y) ≤ g(x ∧ y) + g(x ∨ y) for all
x, y ∈ Rn with the minimum and maximum operators ∧ and ∨ applied component-wise.)
We note that
MTP2 =⇒ CI =⇒ CIS =⇒ Associatedness =⇒ POD.
See Joe (1997), Dhaene et al. (2002a), Denuit et al. (2005), Embrechts et al. (2005),
Kaas et al. (2009), Laeven (2009) and Goovaerts et al. (2011). for further details on these
dependence notions and their connection to VaR and TVaR.
A.2.2 Comparing dependence
The correlation order was introduced in the actuarial literature by Dhaene and Goovaerts
(1996) to find an ordering between random couples X = (X1, X2) and Y = (Y1, Y2) such
that the sums of their components are ordered in the stop-loss (increasing convex order)
sense; see also the concordance order in e.g., Nelsen (2007).
Definition 29 Consider two random couples X = (X1, X2) and Y = (Y1, Y2) inR(F1, F2).
If FX(x1, x2) ≤ FY (x1, x2) for all x ∈ R2, or equivalently, FX(x1, x2) ≤ F Y (x1, x2) for
all x ∈ R2, then X is said to be smaller than Y in the correlation order (denoted by
X ≤corr Y ).
The supermodular order can be seen as a multivariate extension of the correlation
order from two dimensions to higher dimensions, based on supermodular functions.
Definition 30 Let X and Y be two n-dimensional random vectors such that E[g(X)] ≤E[g(Y )] for all supermodular functions g : Rn → R, provided the expectations exist. Then
X is said to be smaller than Y in the supermodular order, denoted by X ≤sm Y .
The multivariate stop-loss order is obtained by substituting the cones of the increasing
convex functions on Rn for the corresponding cone of univariate functions.
32
Definition 31 For two n-dimensional random vectors X and Y , X is said to be smaller
than Y in the multivariate stop-loss order, denoted by X ≤sl Y , if E[g(X)] ≤ E[g(Y )] for
every increasing convex function g : Rn → R.
It is well known that X ≤sl Y if, and only if, g(X) ≤sl g(Y ) for any increasing convex
function g : Rn → R; see Proposition 3.4.65 in Denuit et al. (2005).
Finally, we recall the definition of multivariate usual stochastic order.
Definition 32 For two n-dimensional random vectors X and Y , X is said to be smaller
than Y in the multivariate usual stochastic order, denoted by X ≤st Y , if E[g(X)] ≤E[g(Y )] for every increasing function g : Rn → R.
It is well known that X ≤st Y implies that∑n
i=1Xi ≤st∑n
i=1 Yi and Xi ≤st Yi, for
i = 1, . . . , n. For more detailed properties of the multivariate usual stochastic order,
please refer to Shaked and Shanthikumar (2007).
33